> ## Documentation Index
> Fetch the complete documentation index at: https://docs.profy.cn/llms.txt
> Use this file to discover all available pages before exploring further.

# Cognitive Distillation

> Five-layer cognitive extraction: how to transform human expertise into AI Agent capabilities

Cognitive distillation is Profy's core method for creating Experts. It's not Prompt Engineering — you don't write a role description and then have the AI act it out. Distillation is a **structured deep interview**: the AI asks questions, you answer, and the system extracts the underlying cognitive structures from your responses.

## Five-Layer Cognitive Extraction

The distillation process extracts your expertise across five layers, from "how you understand the world" to "when to say I don't know":

<Steps>
  <Step title="Mental Models">
    The foundational frameworks through which you understand the world. Not knowledge points, but the ways you organize knowledge.

    The extraction standard is triple validation: **cross-domain recurrence** (the same thinking framework appears in 2+ domains), **generative power** (can infer your stance on new problems), and **exclusivity** (not common knowledge everyone shares). Observations passing only 1 validation are downgraded to decision heuristics.
  </Step>

  <Step title="Decision Heuristics">
    How you make choices in specific scenarios. When facing a trade-off between A and B, what's your first instinct? What signals would immediately change your judgment? Under what conditions would you choose "don't act"?
  </Step>

  <Step title="Expression DNA">
    Your unique communication fingerprint — sentence length, analogy density, certainty ratio, transition frequency. Not "imitating how you talk," but quantifying your expression characteristics so the Expert's output style matches yours.
  </Step>

  <Step title="Anti-Pattern Library">
    More important than "knowing what to do" is "knowing what not to do." What kind of advice would you ridicule? Which "common sense" is actually a misconception in your view? What practices are absolutely forbidden in your framework?
  </Step>

  <Step title="Honesty Boundaries">
    A trustworthy Expert must know what it doesn't know. This layer defines scenarios where the Expert should proactively say "I'm not sure" or "I'd suggest consulting another expert."
  </Step>
</Steps>

<Tip>
  Honesty boundaries are the cornerstone of Expert credibility. An AI that claims to know everything is not trustworthy; an AI that clearly knows its boundaries gives more credible advice within its area of expertise.
</Tip>

## The Distillation Process

Distillation is not filling out a form — it's an AI-led deep conversation:

<Steps>
  <Step title="Upload materials">
    Upload materials related to your expertise — articles, speeches, code repositories, conversation logs, even ZIP archives. The distillation AI reads through all materials first to build an initial understanding.
  </Step>

  <Step title="Conversational interview">
    The distillation AI asks you questions. It doesn't ask "what do you do," but rather "when you face the contradiction between A and B, what's the first instinct in your mind." **The AI asks, you answer** — you're not writing a prompt.
  </Step>

  <Step title="Cognitive structuring">
    Five layers of cognitive structure are extracted from the conversation, with each extraction traced back to specific material evidence.
  </Step>

  <Step title="Persona crystallization">
    The structured cognitive profile is transformed into an executable Expert persona — a set of runtime configurations that guide the AI on "how to think."
  </Step>

  <Step title="Scenario calibration">
    The Expert's performance is tested against real scenarios, and deviations are corrected. Every correction is recorded and iteratively optimized. This is the critical step from "roughly correct" to "precisely reproduced."
  </Step>
</Steps>

## Cross-Platform Migration

Already have extensive materials on other platforms? The distillation AI can directly identify and extract from them:

<CardGroup cols={3}>
  <Card title="Project config files" icon="folder">
    Upload .cursor/ directories, AGENTS.md, code convention files to extract coding philosophy and decision preferences.
  </Card>

  <Card title="AI conversation logs" icon="message">
    Import historical conversations with Claude Code, Cursor, and other AIs to extract thinking patterns from your instructions and feedback.
  </Card>

  <Card title="Bulk material packages" icon="file-zipper">
    ZIP-package article collections, note libraries, or knowledge base exports for one-time deep analysis.
  </Card>
</CardGroup>

The distillation AI automatically detects the source platform (OpenClaw / Claude Code / Cursor) and extracts platform-specific configuration semantics accordingly.

## Distillation vs Traditional Methods

| Dimension   | Prompt Engineering                    | Profy Cognitive Distillation                         |
| ----------- | ------------------------------------- | ---------------------------------------------------- |
| Method      | You manually write a role description | AI asks you questions                                |
| Depth       | Surface-level behavior imitation      | Underlying cognitive structures                      |
| Validation  | Subjective judgment                   | Triple validation + scenario calibration             |
| Consistency | Easily "breaks character"             | Derived from mental models, naturally consistent     |
| Evolution   | Manually edit prompts                 | Automatically evolves through usage (Darwin)         |
| Input       | A paragraph of descriptive text       | Materials + deep conversation + scenario calibration |

## Levels and Progression

Distillation quality directly impacts Expert level. Higher quality distillation — more complete mental models, more precise Expression DNA, clearer honesty boundaries — means higher creator levels and better marketplace performance for the Expert.

<Note>
  The goal of cognitive distillation is not to create a "perfect imitation" — but to create a digital agent that inherits core thinking frameworks and can make "the kind of judgment this person would make" in new scenarios. It's an extension of cognition, not a copy of behavior.
</Note>